VT: An Expert Elevator Designer That Uses Knowledge-Based Backtracking

Unlike MOLGEN, VT's decisions about part selection and placement are so interdependent that plausible reasoning (guessing) is a major feature of its search for a solution. Thus, VT's problem-solving strategy is predominantly one of constructing an approximation and successively refining it. Systems that use plausible reasoning must be able to identify bad guesses and improve on these decisions in a way which helps converge on a solution. VT is similar to AIR/CYL (Brown 1985) and PRIDE (Mittal and Araya 1986) in that it uses a knowledge-based approach to direct this search; that is, it uses domain-specific knowledge to decide what past decisions to alter and how to alter them. This approach contrasts with EL (Sussman 1977; Stallman and Sussman 1977), an expert system which shares many architectural features with VT but which uses domain-independent strategies to limit the search during the backtracking phase. As with EL, the VT architecture makes clear the role that domain-specific knowledge plays in the system and the interconnections among decisions used to construct and refine a solution. This architecture provides the basis for VT's explanation facility, which is similar to that of EL and the related CONSTRAINTS language (Sussman and Steele 1980), with some extensions. We have exploited the structure provided by this architecture even further by using it to manage VT's knowledge acquisition. VT's architecture provides structure for a representation of its domain-specific knowledge that reflects the function of the knowledge in problem solving. This representation serves as the basis for an automated knowledge-acquisition tool, SALT (Marcus, n some cases, plausible guessing combined with the ability to backtrack to undo a bad guess can be the most efficient way to solve a problem (Stefik et al. 1983). Even least commitment systems such as MOLGEN (Stefik 1981a, 1981b) are sometimes forced to guess. In the course of designing genetics experiments, MOLGEN tries to avoid making a decision until all constraints that might affect the decision are known. In some cases, this postponement is not possible, and the system becomes stuck; none of the pending decisions can be made with complete confidence. In such a case, a decision based on partial information is needed, and such a decision might be wrong. In this case, a problem solver needs the ability either to backtrack to correct bad decisions or to maintain parallel solutions corresponding to the alternatives at the stuck decision point. However, if alternative guesses exist at each point, and there are many such decision points on each solution path, a commitment to examine every possible combination of alternatives proves unwieldy. Such complexity exists in the VT task domain. VT performs the engineering task of designing elevator systems. It must use the customer's functional specifications to select equipment and produce a parts configuration that meets these specifications as well as safety, installation, and maintenance requirements. Because of the large number of potential part combinations and the need for customizing the layout to the space available in individual buildings, VT must construct a solution. Like MOLGEN, VT tries to order its decisions so that they are made only when all relevant constraints are known; it guesses only when stuck. VT (vertical transportation) is an expert system for handling the design of elevator systems that is currently in use at Westinghouse Elevator Company. Although VT tries to postpone each decision in creating a design until all information that constrains the decision is known, for many decisions this postponement is not possible. In these cases, VT uses the strategy of constructing a plausible approximation and successively refining it. VT uses domain-specific knowledge to guide its backtracking search for successful refinements. The VT architecture provides the basis for a knowledge representation that is used by SALT, an automated knowledge-acquisition tool. SALT was used to build VT and provides an analysis of VT's knowledge base to assess its potential for convergence on a solution.